Harnessing Shallow Features in Pre-Trained Models for Out-of-Distribution Detection

Tong Wei, Bo-Lin Wang, Jiang-Xin Shi, Yu-Feng Li, Min-Ling Zhang

Keywords: out-of-distribution detection, long-tail learning

Abstract: Recognizing out-of-distribution (OOD) samples is essential for deploying robust machine learning systems in the open-world environments. Conventional OOD detection approaches rely on feature representations from the final layer of neuron networks, often neglecting the rich information encapsulated in shallow layers. Leveraging the strengths of transformer-based architectures, we introduce an attention-based fusion module, which dynamically assigns importance weights to representations learned by each Transformer layer and detects OOD samples using the Mahalanobis distance. Compared to existing approaches, our method enables a lightweight fine-tuning of pre-trained models, and retains all feature representations that are beneficial to the OOD detection. We also thoroughly study various parameter-efficient fine-tuning strategies. Our experiments show the benefit of using shallow features, and demonstrate the influence of different Transformer layers. We fine-tune pre-trained models in both class-balanced and long-tailed in-distribution classification tasks, and show that our method achieves state-of-the-art OOD detection performance averaged across nine OOD datasets. The source code is provided in the supplementary material.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3670